From 8d7efe0bde47b784128f4e81c9bec3b39da77e1e Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Mon, 10 Aug 2015 23:05:14 -0500
Subject: [PATCH] Ensure performance-sensitive identifiers are interned
---
loopy/frontend/fortran/translator.py | 7 ++++--
loopy/kernel/__init__.py | 22 ++++++++++++++----
loopy/kernel/creation.py | 33 +++++++++++++++++++--------
loopy/kernel/data.py | 34 +++++++++++++++++++++++++++-
loopy/kernel/tools.py | 2 +-
loopy/tools.py | 10 ++++++++
6 files changed, 91 insertions(+), 17 deletions(-)
diff --git a/loopy/frontend/fortran/translator.py b/loopy/frontend/fortran/translator.py
index 97005cab6..8293035a9 100644
--- a/loopy/frontend/fortran/translator.py
+++ b/loopy/frontend/fortran/translator.py
@@ -25,6 +25,7 @@ THE SOFTWARE.
import re
import six
+from six.moves import intern
import loopy as lp
import numpy as np
@@ -221,7 +222,7 @@ class F2LoopyTranslator(FTreeWalkerBase):
def add_expression_instruction(self, lhs, rhs):
scope = self.scope_stack[-1]
- new_id = "insn%d" % self.insn_id_counter
+ new_id = intern("insn%d" % self.insn_id_counter)
self.insn_id_counter += 1
if self.auto_dependencies and scope.previous_instruction_id:
@@ -447,7 +448,7 @@ class F2LoopyTranslator(FTreeWalkerBase):
def map_IfThen(self, node):
scope = self.scope_stack[-1]
- cond_name = "loopy_cond%d" % self.condition_id_counter
+ cond_name = intern("loopy_cond%d" % self.condition_id_counter)
self.condition_id_counter += 1
assert cond_name not in scope.type_map
@@ -543,6 +544,8 @@ class F2LoopyTranslator(FTreeWalkerBase):
loop_var_suffix += 1
loopy_loop_var = loop_var + "_%d" % loop_var_suffix
+ loopy_loop_var = intern(loopy_loop_var)
+
# }}}
space = isl.Space.create_from_names(isl.DEFAULT_CONTEXT,
diff --git a/loopy/kernel/__init__.py b/loopy/kernel/__init__.py
index 16695e0ad..121a516c1 100644
--- a/loopy/kernel/__init__.py
+++ b/loopy/kernel/__init__.py
@@ -351,7 +351,18 @@ class LoopKernel(RecordWithoutPickling):
for id_str in generate_unique_names(based_on):
if id_str not in used_ids:
- return id_str
+ return intern(id_str)
+
+ def all_group_names(self):
+ result = set()
+ for insn in self.instructions:
+ result.update(insn.groups)
+ result.update(insn.conflicts_with_groups)
+
+ return frozenset(result)
+
+ def get_group_name_generator(self):
+ return _UniqueVarNameGenerator(self.all_group_names())
def get_var_descriptor(self, name):
try:
@@ -577,7 +588,8 @@ class LoopKernel(RecordWithoutPickling):
def all_inames(self):
result = set()
for dom in self.domains:
- result.update(dom.get_var_names(dim_type.set))
+ result.update(
+ intern(n) for n in dom.get_var_names(dim_type.set))
return frozenset(result)
@memoize_method
@@ -588,7 +600,8 @@ class LoopKernel(RecordWithoutPickling):
for dom in self.domains:
result.update(set(dom.get_var_names(dim_type.param)) - all_inames)
- return frozenset(result)
+ from loopy.tools import intern_frozenset_of_ids
+ return intern_frozenset_of_ids(result)
def outer_params(self, domains=None):
if domains is None:
@@ -600,7 +613,8 @@ class LoopKernel(RecordWithoutPickling):
all_inames.update(dom.get_var_names(dim_type.set))
all_params.update(dom.get_var_names(dim_type.param))
- return all_params-all_inames
+ from loopy.tools import intern_frozenset_of_ids
+ return intern_frozenset_of_ids(all_params-all_inames)
@memoize_method
def all_insn_inames(self):
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index e891f0626..556bc3a74 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -26,6 +26,7 @@ THE SOFTWARE.
import numpy as np
+from loopy.tools import intern_frozenset_of_ids
from loopy.symbolic import IdentityMapper, WalkMapper
from loopy.kernel.data import (
InstructionBase, ExpressionInstruction, SubstitutionRule)
@@ -33,7 +34,7 @@ import islpy as isl
from islpy import dim_type
import six
-from six.moves import range, zip
+from six.moves import range, zip, intern
import re
import sys
@@ -216,7 +217,7 @@ def parse_insn(insn):
opt_value = option[equal_idx+1:].strip()
if opt_key == "id":
- insn_id = opt_value
+ insn_id = intern(opt_value)
elif opt_key == "id_prefix":
insn_id = UniqueName(opt_value)
elif opt_key == "priority":
@@ -235,17 +236,18 @@ def parse_insn(insn):
insn_deps_is_final = True
opt_value = (opt_value[1:]).strip()
- insn_deps = frozenset(dep.strip() for dep in opt_value.split(":")
+ insn_deps = frozenset(
+ intern(dep.strip()) for dep in opt_value.split(":")
if dep.strip())
elif opt_key == "groups":
insn_groups = frozenset(
- grp.strip() for grp in opt_value.split(":")
+ intern(grp.strip()) for grp in opt_value.split(":")
if grp.strip())
elif opt_key == "conflicts":
conflicts_with_groups = frozenset(
- grp.strip() for grp in opt_value.split(":")
+ intern(grp.strip()) for grp in opt_value.split(":")
if grp.strip())
elif opt_key == "inames":
@@ -255,10 +257,10 @@ def parse_insn(insn):
else:
forced_iname_deps_is_final = True
- forced_iname_deps = frozenset(opt_value.split(":"))
+ forced_iname_deps = intern_frozenset_of_ids(opt_value.split(":"))
elif opt_key == "if":
- predicates = frozenset(opt_value.split(":"))
+ predicates = intern_frozenset_of_ids(opt_value.split(":"))
elif opt_key == "tags":
tags = tuple(
@@ -284,7 +286,10 @@ def parse_insn(insn):
"be variable or subscript" % lhs)
return ExpressionInstruction(
- id=insn_id,
+ id=(
+ intern(insn_id)
+ if not isinstance(insn_id, (type(None), UniqueName))
+ else None),
insn_deps=insn_deps,
insn_deps_is_final=insn_deps_is_final,
groups=insn_groups,
@@ -326,7 +331,17 @@ def parse_insn(insn):
def parse_if_necessary(insn, defines):
if isinstance(insn, InstructionBase):
- yield insn, []
+ yield insn.copy(
+ id=intern(insn.id) if insn.id is not None else None,
+ insn_deps=frozenset(intern(dep) for dep in insn.insn_deps),
+ groups=frozenset(intern(grp) for grp in insn.groups),
+ conflicts_with_groups=frozenset(
+ intern(grp) for grp in insn.conflicts_with_groups),
+ forced_iname_deps=frozenset(
+ intern(iname) for iname in insn.forced_iname_deps),
+ predicates=frozenset(
+ intern(pred) for pred in insn.predicates),
+ ), []
return
elif not isinstance(insn, str):
raise TypeError("Instructions must be either an Instruction "
diff --git a/loopy/kernel/data.py b/loopy/kernel/data.py
index b88929358..2d9faa4cf 100644
--- a/loopy/kernel/data.py
+++ b/loopy/kernel/data.py
@@ -25,6 +25,7 @@ THE SOFTWARE.
"""
+from six.moves import intern
import numpy as np
from pytools import Record, memoize_method
from loopy.kernel.array import ArrayBase
@@ -185,6 +186,8 @@ def parse_tag(tag):
class KernelArgument(Record):
def __init__(self, **kwargs):
+ kwargs["name"] = intern(kwargs.pop("name"))
+
dtype = kwargs.pop("dtype", None)
if isinstance(dtype, np.dtype):
@@ -339,7 +342,7 @@ class TemporaryVariable(ArrayBase):
if base_indices is None:
base_indices = (0,) * len(shape)
- ArrayBase.__init__(self, name=name,
+ ArrayBase.__init__(self, name=intern(name),
dtype=dtype, shape=shape,
dim_tags=dim_tags, order="C",
base_indices=base_indices, is_local=is_local,
@@ -512,6 +515,9 @@ class InstructionBase(Record):
forced_iname_deps_is_final, forced_iname_deps, priority,
boostable, boostable_into, predicates, tags):
+ if insn_deps is None:
+ insn_deps = frozenset()
+
if groups is None:
groups = frozenset()
@@ -531,6 +537,17 @@ class InstructionBase(Record):
if tags is None:
tags = ()
+ # Periodically reenable these and run the tests to ensure all
+ # performance-relevant identifiers are interned.
+ #
+ # from loopy.tools import is_interned
+ # assert is_interned(id)
+ # assert all(is_interned(dep) for dep in insn_deps)
+ # assert all(is_interned(grp) for grp in groups)
+ # assert all(is_interned(grp) for grp in conflicts_with_groups)
+ # assert all(is_interned(iname) for iname in forced_iname_deps)
+ # assert all(is_interned(pred) for pred in predicates)
+
assert isinstance(forced_iname_deps, frozenset)
assert isinstance(insn_deps, frozenset) or insn_deps is None
assert isinstance(groups, frozenset)
@@ -650,6 +667,21 @@ class InstructionBase(Record):
# }}}
+ def __setstate__(self, val):
+ super(InstructionBase, self).__setstate__(val)
+
+ from loopy.tools import intern_frozenset_of_ids
+
+ self.id = intern(self.id)
+ self.insn_deps = intern_frozenset_of_ids(self.insn_deps)
+ self.groups = intern_frozenset_of_ids(self.groups)
+ self.conflicts_with_groups = (
+ intern_frozenset_of_ids(self.conflicts_with_groups))
+ self.forced_iname_deps = (
+ intern_frozenset_of_ids(self.forced_iname_deps))
+ self.predicates = (
+ intern_frozenset_of_ids(self.predicates))
+
# }}}
diff --git a/loopy/kernel/tools.py b/loopy/kernel/tools.py
index b59c40731..e811935bb 100644
--- a/loopy/kernel/tools.py
+++ b/loopy/kernel/tools.py
@@ -204,7 +204,7 @@ def find_all_insn_inames(kernel):
# current inames refer to.
if par in kernel.all_inames():
- inames_new.add(par)
+ inames_new.add(intern(par))
# If something writes the bounds of a loop in which I'm
# sitting, I had better be in the inames that the writer is
diff --git a/loopy/tools.py b/loopy/tools.py
index e734417d6..861d15568 100644
--- a/loopy/tools.py
+++ b/loopy/tools.py
@@ -30,6 +30,7 @@ from loopy.symbolic import WalkMapper as LoopyWalkMapper
from pymbolic.mapper.persistent_hash import (
PersistentHashWalkMapper as PersistentHashWalkMapperBase)
import six # noqa
+from six.moves import intern
if six.PY2:
@@ -216,4 +217,13 @@ def remove_common_indentation(code, require_leading_newline=True,
# }}}
+
+def is_interned(s):
+ return s is None or intern(s) is s
+
+
+def intern_frozenset_of_ids(fs):
+ return frozenset(intern(s) for s in fs)
+
+
# vim: foldmethod=marker
--
GitLab